Saif 4B (Gyr Technology)v1.7 · 10 s rule
Data: JevBench v1.8.1, published 2026-10-30 · method v1.7 (10 s rule), measured under the rule
Capability 66.4 · Not ranked in the current release
Measured on the 27 Oct 2026 draw, which is not yet equated to the reference scale; stronger anchor runs for that draw are being measured. Shown unranked until then.
Last measured: 2026-10-29 · measurement release: v1.8.1
CC-BY-NC-4.0 (LICENSE file) - non-commercial
intelligence
43.0
calibration
89.9
speed
91.5
cost
81.2
Composite (secondary)
51.3 · Not ranked on this board
Cost and measurement conditions
$0.0042 per 1,000 decisions (estimated)
GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 643,766 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00424838271972791/1,000; old list-price estimate USD 0.012970436534587/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00424838271972791/1,000. Estimated, not a bill. USD 0.00424838272 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00682261406 over all 1,500 items).
p50 latency: 0.2 s
Self-hosted latency estimate under the shared method (assumption, not measured)
Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition (97,887 MiB reported, driver 580.178.04), FIN-03 (Finland, EU), instance type 1RTXPRO6000.30V, image 24.04.cuda13.0. Measured 29 Oct 2026. Author's native HTTP server on loopback, pinned container and local hash-verified weights, read-only mounts, private output, empty credentials, offline flags and network cut before sealed input. Weights-verify receipt reports 35 files verified; actual runtime distribution, author-source and base/model tokenizer hash receipts were retained. Public smoke passed before sealed input: timed probes at 256/2,048/8,192/16,384/32,768/49,152/82,000 tokens all returned HTTP 200 with usage equal to offline counts after separately recorded warm-ups. The maximum timed 82,000-token noul probe was 8.645 s; a separate long-choice probe took 8.642 s. The 131,073-token context-refusal probe and invalid one-level score probe returned native HTTP 422. Frozen-adapter public noul/choice/score/255-choice/82,000-token-state/Unicode-JSON probes passed with native usage equality. TTL-fit admitted with 3,513.56 s remaining against 1,500 s required, forecasting 0 items over 10 s; this forecast was an admission estimate, not a scoring exclusion. The scored server passed a separate public pre-warm. Complete 1,500-row run; scored loop 113.2420175075531 s. Raw output sha256 1a7b04e171d8c23f88927b24810563618604f5d64bc2bfa4b333121aba1db243 (independently verified sealed artifact). Row-completion reports complete, teardown_confirmed, node_shred_verified and local_shred_verified true; node-shred and local-shred receipts verified, and teardown receipt confirms the exact owned VM closed. GPU rental is separate from the benchmark's estimated model tariff.
Measured latency over 1,500 decisions: p50 0.043 s · p95 0.076 s
0 decisions over 10 s.
A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.